You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: scenarios/security_arena/docs/phase2.md
+10Lines changed: 10 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -220,6 +220,16 @@ The defender is **stateless** — each round is a fresh conversation. It sees:
220
220
- Make security instructions scenario-aware. Parse `defender_task` and `defender_role` to generate tailored defenses.
221
221
- Stay helpful — a defender that refuses everything fails the normal user test and won't appear on the leaderboard.
222
222
223
+
### Example submission
224
+
225
+
See this PR for a complete example of what a Phase 2 submission looks like: [**PR #34: Add reasoning to attacker, two-pass defense to defender**](https://github.com/LambdaLabsML/agentbeats-lambda/pull/34)
226
+
227
+
It shows:
228
+
- Adding a reasoning step to the attacker (`agents/attacker/agent.py`)
229
+
- Adding a two-pass defense to the defender (`agents/defender/agent.py`)
230
+
- Only files inside `agents/attacker/` and `agents/defender/` are modified — no framework changes needed
231
+
- The commit message uses `[submit]` to trigger the submission workflow
0 commit comments